Build · 5 Jun 2026

Octavo fits in 16 MB, with a weaker guarantee

One false accept in 4,096, against one in 16.7 million on Quarto.

Build
4c81e0a2
Budget
16 MB
Fingerprint
12 bits
False accept
1 in 4,096

Specifics

The numbers, and where they come from

Figures on this page, with the basis of each
QuantityValueBasis
Budget16 MB XIPStated
Fingerprint width12 bitsStated
False accept rate1 in 4,096Computed
Quarto, for comparison1 in 16.7MComputed
Octavo, 12 bits1 in 4,096Quarto, 24 bits1 in 16.7MFolio, full widthnot binding
Inverse false accept rate by format, logarithmic. Even on a scale that flatters it, Octavo is visibly the odd one out.

The result

It fits, and the fit costs something

Build 4c81e0a2 is the first artefact to run entirely from 16 MB of execute-in-place flash with no card attached. That is the hardware floor most cheap embedded targets can actually meet, and reaching it was the point of the format.

Twelve-bit fingerprints are all the remaining budget allows, which puts the false accept rate at one in 4,096 against one in 16.7 million on Quarto. That is four orders of magnitude, and it is a weaker guarantee rather than merely less content.

How we talk about it

In the same breath as the name

The rule is that Octavo’s weaker guarantee appears in the same sentence as the word Octavo, every time, including in sales conversations. It does not go in a footnote, an appendix or a specification annex, because a limit found in a footnote reads as a limit that was discovered rather than one that was known.

The three formats are not a capability ladder. Inference cost here is effectively zero, so the pricing logic that justifies a model size ladder does not exist, and presenting them as tiers would import an argument that does not apply.

Where Octavo is not appropriate

If your deployment sees more than a few million lookups per device lifetime, size for Quarto. The arithmetic for that decision is published rather than gatekept.